{"id":"b82e0c4f-d48b-4ae1-b30e-6c30a4549bfa","arxiv_id":"2608.02145","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"UniqueSplat conditions a feed-forward 3D Gaussian Splatting predictor on the target view through a two-branch hypernetwork, improving sparse-view novel view synthesis on standard benchmarks.","lead":"This paper adds a view-conditioned hypernetwork to MVSplat, a feed-forward 3D Gaussian Splatting model, so the network weights shift with each target viewpoint and produce customized Gaussians. On RealEstate10K, ACID and DTU it reports small but consistent gains over prior sparse-view NVS methods, including across datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim is under-attributed: Table IV ablates branches but not capacity or target-view information, so the reported gains may come from 26.1M added parameters rather than per-view Gaussian customization.","rationale":"The reader's weakest assumption is the same concern I identified: the ablations do not control for added capacity or for the presence of target-view information, so the central claim that view-conditioning causes the improvement remains conditional. This is a genuine load-bearing issue because the paper's novelty, stated in the abstract, contributions, and Section I, is specifically the dynamically adjusted, view-conditioned Gaussian prediction. If a capacity-matched static predictor achieves the same PSNR, then the method reduces to a larger MVSplat variant and the central claim collapses. If the constant-embedding variant performs equally well, then the view-specific branch is not actually using the target view in a meaningful way. Both are concrete, testable failure modes. I do not see an internal inconsistency or a reason to reject the paper outright; the method is clearly described, benchmarked on three datasets, and the reported gains are consistent. But the current experimental design cannot distinguish mechanism from capacity. A 2x2 ablation with matched parameters and a constant-embedding control would settle the attribution, which is exactly the condition under which the paper should be accepted. I therefore agree with the reader's CONDITIONAL verdict and would not move it.","tokens_in":18924,"tokens_out":4825,"duration_ms":49967,"concrete_test":"Run a 2x2 controlled ablation on RealEstate10K: (A) a static predictor with the full model's 38.1M parameters, e.g., widen MVSplat's U-Net/transformer or add fixed input-independent hypernetwork-generated weights, trained with the same loss, schedule, and MVSplat initialization; (B) the full UniqueSplat with the view-specific embedding e_s replaced by a constant vector shared across all target views, so no target-view projection enters but capacity is unchanged. If (A) matches or exceeds 27.28 dB PSNR, the gain is capacity, not conditioning. If (B) does not drop below (A), the per-view weight prediction is not the operative mechanism. Report paired per-scene bootstrap confidence intervals for PSNR on all arms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the abstract and Section I, is that per-view Gaussian customization causes the reported improvements over MVSplat and TranSplat. The evidence for this attribution is incomplete. Table IV compares only three configurations: the 12.0M MVSplat primary network, that network plus a view-agnostic branch, and the full 38.1M model. The view-agnostic branch generates a fixed weight tensor from a learned embedding; replacing predictor weights with W_a = H_a(e_a) is effectively a reparameterization with added capacity, not a demonstrated mechanism for extracting shareable knowledge. The view-specific branch adds a ResNet18 backbone, target-view projection, correlation fusion, and an MLP generating ΔW, so most of the 26.1M extra parameters are uncontrolled. No static baseline with capacity matched to the full model is reported, and no ablation removes target-view information while holding capacity fixed. Table III shows the parameter gap directly: MVSplat 12.0M versus UniqueSplat 38.1M. The cross-dataset results in Table II are subject to the same confound. The view-agnostic embedding is a learned constant from the training distribution and cannot itself adapt to DTU, while the view-specific branch is untested against a capacity-matched un-conditioned predictor. The reported numbers may therefore reflect additional parameters and the MVSplat initialization rather than the stated view-conditioning mechanism. The paper is otherwise coherent and the gains are plausible, but the causal attribution is not established by the current experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniqueSplat, a feed-forward 3D Gaussian Splatting method that predicts view-dependent Gaussians by injecting target-view information into the predictor's weights. The model consists of a primary network (based on MVSplat) and a two-branch hypernetwork: a view-agnostic branch that produces a fixed weight tensor from a learned constant embedding, and a view-specific branch that extracts features from input views, projects them to the target view via epipolar geometry, and generates a weight update. The fused weights are applied to the convolutional layers of the predictor, which then estimates Gaussian parameters for rendering the target view. Experiments on RealEstate10K, ACID, and DTU report state-of-the-art results in both intra-dataset and cross-dataset settings, along with an efficiency comparison and ablations of the two branches.","tokens_in":19197,"tokens_out":5707,"duration_ms":54536,"significance":"If the proposed mechanism is responsible for the reported gains, the contribution is significant: UniqueSplat is among the first feed-forward 3DGS models to condition the Gaussian predictor on the query view, and the cross-dataset gains on DTU are large. The paper also reports a favorable efficiency trade-off relative to TranSplat (38.1M parameters vs. 110.5M) and includes failure cases and qualitative comparisons. However, the central attribution claim is not yet established. The ablations do not control for the large increase in parameter count or for the MVSplat initialization, and the view-agnostic branch is a fixed weight tensor rather than a dynamic sharing mechanism. With the missing capacity-matched and simple-conditioning baselines, the reported improvements could plausibly come from added model capacity rather than from view-conditioned Gaussian customization.","major_comments":[{"comment":"The ablation design does not isolate the proposed view-conditioning mechanism. The full model (38.1M) is compared only against MVSplat (12.0M) and MVSplat plus the view-agnostic branch; the view-specific branch adds a ResNet18 encoder, a projection module, and a hypernetwork, so the 0.89 dB gain over MVSplat on RealEstate10K could be caused by additional capacity rather than by per-view conditioning. Please add (i) a capacity-matched static predictor with the same parameter count and the same MVSplat initialization, (ii) a variant that removes target-view information from the view-specific branch while keeping the architecture fixed (for example, replacing the target projection matrices with a fixed reference projection), and (iii) a simple conditioning baseline such as concatenating the target pose to the cost volume or using FiLM conditioning. Without these controls, the central claim that view-conditioning rather than capacity drives the improvements is not supported.","section":"Section IV-E, Table IV"},{"comment":"The view-agnostic branch is described as extracting shareable information from diverse input views, but at inference it computes W_a = H_a(e_a) once from a single learned constant e_a that does not depend on the input images, the scene, or the target view. This is a fixed weight tensor, and adding it is essentially a reparameterization of the predictor's initial weights. The gains from variant (a) to variant (b) in Table IV (0.31 dB on RealEstate10K and 0.45 dB on ACID) may therefore reflect added capacity or the MVSplat initialization rather than any learned cross-view abstraction. Please clarify the intended interpretation and provide an ablation where e_a is conditioned on the input scene, or explicitly state that this branch serves as a learned initialization.","section":"Section III-B1, Eq. (1), and Algorithm 1"},{"comment":"The cross-dataset claim is subject to the same confound and to an additional transfer concern. The view-agnostic embedding is a constant learned on RealEstate10K and is not updated at test time, so it cannot provide scene- or dataset-specific adaptation on ACID or DTU; the only adaptive component is the view-specific branch. Given the large DTU improvement (16.01 vs. 13.94 for MVSplat), please report per-scene standard deviations or error bars and analyze whether the gain is attributable to target-view conditioning by including the capacity-matched and simple-conditioning baselines in the cross-dataset setting. Without this evidence, the generalization claim is not independently established.","section":"Section IV-B, Table II"}],"minor_comments":[{"comment":"Table I has the header typo 'LPISP' instead of 'LPIPS'; Figure 4's caption contains 'Qualitatitive' and 'corss-dataset'; Table II's caption says 'Qualitative comparison' even though the table is quantitative. These should be corrected.","section":"Tables and captions"},{"comment":"The correlation function C_ij(M1, M2) is not precisely defined: the text writes (M_i^1, M_j^2) / (||M_i^1|| ||M_j^2||), but it is unclear whether the superscripts index the two matrices or whether M_i^1 means the i-th row of M1. Please define the indexing explicitly.","section":"Section III-B2, Eq. (10)"},{"comment":"The notation is inconsistent: the text says the hypernetwork H_s produces a weight matrix W_s, but Eq. (12) defines ΔW = H_s(e_s) and Eq. (13) writes W_H = W + ΔW. Please align the notation (e.g., set W_s = ΔW) and specify which layers of the predictor are affected.","section":"Section III-B2, Eqs. (12)-(13)"},{"comment":"Table III mixes settings: the ACID column reproduces the intra-dataset numbers from Table I, while the DTU column reproduces the cross-dataset numbers from Table II. Since the table is intended as an efficiency comparison, please state explicitly which training/test protocol is used for each dataset, or report the two settings separately.","section":"Section IV-D, Table III"}],"recommendation":"major_revision","confidential_remarks":"I see no grounds for rejection: the paper is coherent, the experiments are extensive, and the mechanism is plausible. The main issue is that the experimental design does not yet prove the central attribution claim. If the authors add the requested capacity-matched and simple-conditioning baselines, clarify the role of the view-agnostic branch, and fix the table inconsistencies, the contribution could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental extension of MVSplat, and the architecture is genuinely new. But the central claim—that per-view Gaussian customization causes the gains—is not yet supported because the full model has 3x the parameters of MVSplat and there is no capacity-matched baseline.\n\nWhat's genuinely new: the two-branch view-conditioned hypernetwork, where a view-agnostic branch generates base weights and a view-specific branch generates a delta from a target-view projection and correlation of input features. That goes beyond fixed-Gaussian predictors and prior hypernetwork NVS methods. The experiments are broad: three datasets, intra- and cross-dataset, with consistent improvements (e.g., +0.89 dB PSNR over MVSplat on RealEstate10K, +0.81 on ACID, and +2.07 on DTU cross-dataset). The ablations show each branch contributes, and the efficiency table is useful.\n\nSoft spots: the attribution problem is real. Table IV compares the 12M-parameter MVSplat against variants with growing parameter counts, ending at 38.1M. The view-specific branch alone adds a ResNet18 backbone and projection MLPs. There's no control that adds comparable capacity without view conditioning, nor a simple conditioning baseline (e.g., concatenating the target pose into the cost volume). So the observed gains could partly be capacity, not the hypothesized mechanism. Also, no error bars or significance tests are reported—these are single-run averages. The cross-dataset claim rests on the same confound; the view-agnostic embedding is a learned constant, so it's not obvious why it transfers to DTU except through added model capacity. Minor: Table II is labeled 'Qualitative' but reports numbers, and there's a typo 'LISP' in Table I.\n\nBottom line: this is a well-executed paper with a plausible idea, and the results are probably real. But the load-bearing claim about view-conditioning needs a capacity-matched ablation before it's convincing. I'd send it to review and ask for that experiment, plus code or a detailed configuration.","headline":"Useful incremental extension of MVSplat with a genuinely new two-branch view-conditioned hypernetwork; the capacity confound leaves the central attribution unproven.","tokens_in":19792,"tokens_out":2072,"would_cite":true,"duration_ms":19030,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that conditioning the Gaussian predictor on the target view via a two-branch hypernetwork gives sharper sparse-view renderings and stronger cross-dataset generalization than fixed-Gaussian feed-forward models.","keywords":["3D Gaussian Splatting","view-conditioned hypernetwork","generalizable novel view synthesis","sparse-view 3D reconstruction","feed-forward reconstruction","cross-dataset generalization","cost volume","dynamic Gaussian prediction"],"falsifier":"Train two variants of the same primary network: one with fixed weights and the same added parameter count, and one that concatenates the target camera pose into the encoder features instead of generating weights; if either matches or exceeds the reported PSNR on RealEstate10K, the claim that view-customized weights cause the improvement is falsified. Alternatively, check whether the Gaussians predicted for two different target views of the same scene differ by more than floating-point noise; if they are effectively identical, the view-specific branch is inert.","tokens_in":18645,"feed_emoji":"🎥","tokens_out":9433,"duration_ms":75875,"temperature":0.7,"pith_summary":"Feed-forward 3D Gaussian Splatting methods currently predict one fixed set of Gaussians per scene, tuned to render every viewpoint about equally well. UniqueSplat claims that this leaves quality on the table, because the target viewpoint is known at inference time but never used when the Gaussians are predicted. The paper proposes conditioning the Gaussian predictor on the query view through a two-branch hypernetwork: one branch supplies stable, shareable weights, and the other injects target-view-specific adjustments. If it is right, a single forward pass renders sharper novel views than fixed-Gaussian baselines and transfers better to unseen datasets. The reported gains include PSNR increases from 26.39 to 27.28 on RealEstate10K and from 13.94 to 16.01 on the cross-dataset DTU test.","feed_headline":"View-conditioned Gaussians beat fixed-view splatting in sparse views","feed_subtitle":"A two-branch hypernetwork injects target-view knowledge into the predictor, improving both quality and cross-dataset transfer.","key_machinery":"The central object is the view-conditioned hypernetwork, a weight generator for the Gaussian predictor. A view-agnostic branch produces a base weight $W=H_a(e_a)$ from a learned embedding $e_a$ that abstracts knowledge shared across views, while a view-specific branch produces an adjustment $\\Delta W=H_s(e_s)$ from an embedding $e_s$ obtained by projecting source features onto the target view along epipolar lines. The fused weight $W_H = W + \\Delta W$ is reshaped and injected into the predictor's convolutional layers, so the predicted Gaussian parameters $\\mu$, $\\Sigma$, $\\alpha$ and $c$ depend on the query view and are rendered by rasterization-based splatting.","core_discovery":"On its own terms, the paper's discovery is that the network which turns multi-view features into 3D Gaussian parameters does not have to be a fixed function shared across all target views. By generating the predictor's convolutional weights from a learned view-agnostic embedding plus a view-specific embedding derived from the target camera and the source features, the model produces a different set of Gaussians for each query view while keeping training single-stage and feed-forward. The authors argue that this dynamic-weight mechanism removes the cracks, blurs and artifacts visible in fixed-Gaussian baselines, and they support it with consistent PSNR, SSIM and LPIPS gains in both intra-dataset and cross-dataset settings. They also report that the added machinery is lighter than the transformer-based alternative: 38.1M parameters, 0.065s and 3.12GB versus 110.5M parameters, 0.083s and 4.98GB.","pith_inferences":["I infer that the view-agnostic branch probably acts as a learned initialization that anchors the fused weights near a stable average, while the view-specific branch supplies a correction; measuring the norm of $\\Delta W$ across target views would show how much of the weight space is actually customized.","I infer the same conditioning principle could extend beyond Gaussian splatting to depth or cost-volume predictors, where the target camera would reshape the cost volume or the refinement head.","I infer the cross-dataset gain on object-centric scenes may come partly from the epipolar feature projection rather than from the hypernetwork itself, since that projection injects explicit geometric consistency; an ablation separating projection from weight modulation would settle it.","A testable extension is to condition on per-pixel ray directions rather than a single global target embedding, which would allow spatially varying customization of the Gaussian field."],"forward_implications":["If the claim holds, feed-forward sparse-view reconstruction no longer has to compromise for all viewpoints at once: each target query can receive its own Gaussian field in a single forward pass.","The same two-branch weight-generation scheme could be attached to other cost-volume or epipolar predictors, not only the Gaussian head demonstrated here.","Cross-dataset transfer improves: training once on one distribution yields stronger zero-shot rendering on other datasets than fixed-Gaussian baselines, suggesting view conditioning acts as a partial domain-invariant prior.","The overhead stays practical: the reported parameter count, latency and memory sit between the lightweight baseline and the transformer-based alternative, so the specialization does not require per-scene optimization."],"supporting_citations":[{"why":"MVSplat supplies the primary network architecture and the fixed-Gaussian baseline that UniqueSplat extends and compares against.","marker":"[1]"},{"why":"pixelSplat defines the feed-forward Gaussian-splatting formulation built on epipolar transformers that the paper contrasts with view-conditioning.","marker":"[20]"},{"why":"TranSplat is the strongest recent transformer-based baseline in the intra- and cross-dataset comparisons.","marker":"[18]"},{"why":"RealEstate10K is the main training and intra-test benchmark for all central comparisons.","marker":"[65]"},{"why":"ACID provides the second intra-test benchmark and a cross-dataset generalization target.","marker":"[66]"},{"why":"DTU is the object-centric zero-shot target on which the largest reported generalization gains appear.","marker":"[68]"},{"why":"Hypernetworks establishes the weight-generation mechanism that the two-branch design adapts for view conditioning.","marker":"[55]"},{"why":"3D Gaussian Splatting defines the primitive representation and rasterization renderer that the predictor outputs and renders.","marker":"[14]"}],"fun_headline_variants":["UniqueSplat: Hypernetwork tunes Gaussians per view","View-conditioned splatting via dynamic network weights","Two-branch hypernetwork adapts Gaussians to each view","Feed-forward 3D with per-view Gaussian prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper attributes the performance gains to the view-conditioning mechanism, but the ablations compare only the presence of the two branches and do not test a static predictor with matched capacity or a simpler injection of the target pose, so the gains could in principle come from the extra parameters rather than from conditioning.","fun_headline_variants_meta":{"raw":{"variants":["UniqueSplat: Hypernetwork tunes Gaussians per view","View-conditioned splatting via dynamic network weights","Two-branch hypernetwork adapts Gaussians to each view","Feed-forward 3D with per-view Gaussian prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1579,"prompt_tokens":973,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":540}},"tokens_in":589,"tokens_out":606,"duration_ms":5737,"temperature":1.0,"reasoning_tokens":540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:01:35.832039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two variants of the same primary network: one with fixed weights and the same added parameter count, and one that concatenates the target camera pose into the encoder features instead of generating weights; if either matches or exceeds the reported PSNR on RealEstate10K, the claim that view-customized weights cause the improvement is falsified. Alternatively, check whether the Gaussians predicted for two different target views of the same scene differ by more than floating-point noise; if they are effectively identical, the view-specific branch is inert.","supporting_citations":[{"cited_title":"Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,","cited_arxiv_id":null,"evidence_quote":"MVSplat supplies the primary network architecture and the fixed-Gaussian baseline that UniqueSplat extends and compares against."},{"cited_title":"Large scale multi-view stereopsis evaluation,","cited_arxiv_id":null,"evidence_quote":"DTU is the object-centric zero-shot target on which the largest reported generalization gains appear."}],"review_version":2}